CIAStereo: Stereo Matching Network With Cross-It-Eration-Aligned Disparity Regularization

Recent advances in iterative networks have improved stereo matching accuracy but suffer from slow inference and high computational costs. We propose the CIAStereo method, which enables us to achieve swift inference speeds while maintaining competitive accuracy. We design an efficient SRExtractor and propose channel-wise feature alignment to improve the accuracy. Additionally, we enhance iteration efficiency of the iterative unit by introducing cross-iteration disparity regularization. We achieve 92.5 % faster feature extraction, 17.1 % faster inference, and 1.2 % higher accuracy on KITTI-2015 (D1-bg) compared to baseline model, using only <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 6. 8 \%}$</tex> of its parameters.